An Ensemble Classification Method Based on Input Clustering and Classifiers Expected Reliability

Abstract

In this paper a novel ensemble method (EM) for classification tasks is described. The proposed approach is based on the use of a set of classifiers, each of which is trained by exploiting a different subset of the available training data, which are created by partitioning the input space by means of a self organizing map (SOM) based clustering algorithm. Subsequently, the reliability of each classifier belonging to the ensemble is measured according to the classification accuracy on whole dataset and each classifier is associated to a feed forward neural network, which is able to self-estimate the reliability of single classifiers when coping with a new data. The estimated reliabilities are used in the ensemble aggregation phase in order to provide the final classification of new patterns. The method, tested on literature datasets coming from the UCI repository, achieved satisfactory results improving the classification accuracy with respect to other popular ensemble techniques.